grpc 1.28.0 → 1.37.0

data/third_party/boringssl-with-bazel/src/crypto/fipsmodule/rand/urandom.c

@@ -65,38 +65,12 @@
 #include <openssl/thread.h>
 #include <openssl/mem.h>
 
+#include "getrandom_fillin.h"
 #include "../delocate.h"
 #include "../../internal.h"
 
 
-#if defined(OPENSSL_LINUX)
-
-#if defined(OPENSSL_X86_64)
-#define EXPECTED_NR_getrandom 318
-#elif defined(OPENSSL_X86)
-#define EXPECTED_NR_getrandom 355
-#elif defined(OPENSSL_AARCH64)
-#define EXPECTED_NR_getrandom 278
-#elif defined(OPENSSL_ARM)
-#define EXPECTED_NR_getrandom 384
-#elif defined(OPENSSL_PPC64LE)
-#define EXPECTED_NR_getrandom 359
-#endif
-
-#if defined(EXPECTED_NR_getrandom)
-#define USE_NR_getrandom
-
-#if defined(__NR_getrandom)
-
-#if __NR_getrandom != EXPECTED_NR_getrandom
-#error "system call number for getrandom is not the expected value"
-#endif
-
-#else  // __NR_getrandom
-
-#define __NR_getrandom EXPECTED_NR_getrandom
-
-#endif  // __NR_getrandom
+#if defined(USE_NR_getrandom)
 
 #if defined(OPENSSL_MSAN)
 void __msan_unpoison(void *, size_t);
@@ -119,28 +93,12 @@ static ssize_t boringssl_getrandom(void *buf, size_t buf_len, unsigned flags) {
   return ret;
 }
 
-#endif  // EXPECTED_NR_getrandom
-
-#if !defined(GRND_NONBLOCK)
-#define GRND_NONBLOCK 1
-#endif
-#if !defined(GRND_RANDOM)
-#define GRND_RANDOM 2
-#endif
-
-#endif  // OPENSSL_LINUX
-
-// rand_lock is used to protect the |*_requested| variables.
-DEFINE_STATIC_MUTEX(rand_lock)
+#endif  // USE_NR_getrandom
 
-// The following constants are magic values of |urandom_fd|.
-static const int kUnset = 0;
+// kHaveGetrandom in |urandom_fd| signals that |getrandom| or |getentropy| is
+// available and should be used instead.
 static const int kHaveGetrandom = -3;
 
-// urandom_fd_requested is set by |RAND_set_urandom_fd|. It's protected by
-// |rand_lock|.
-DEFINE_BSS_GET(int, urandom_fd_requested)
-
 // urandom_fd is a file descriptor to /dev/urandom. It's protected by |once|.
 DEFINE_BSS_GET(int, urandom_fd)
 
@@ -179,14 +137,9 @@ static void maybe_set_extra_getrandom_flags(void) {
 DEFINE_STATIC_ONCE(rand_once)
 
 // init_once initializes the state of this module to values previously
-// requested. This is the only function that modifies |urandom_fd| and
-// |urandom_buffering|, whose values may be read safely after calling the
-// once.
+// requested. This is the only function that modifies |urandom_fd|, which may be
+// read safely after calling the once.
 static void init_once(void) {
-  CRYPTO_STATIC_MUTEX_lock_read(rand_lock_bss_get());
-  int fd = *urandom_fd_requested_bss_get();
-  CRYPTO_STATIC_MUTEX_unlock_read(rand_lock_bss_get());
-
 #if defined(USE_NR_getrandom)
   int have_getrandom;
   uint8_t dummy;
@@ -229,31 +182,16 @@ static void init_once(void) {
   abort();
 #endif
 
-  if (fd == kUnset) {
-    do {
-      fd = open("/dev/urandom", O_RDONLY);
-    } while (fd == -1 && errno == EINTR);
-  }
+  int fd;
+  do {
+    fd = open("/dev/urandom", O_RDONLY);
+  } while (fd == -1 && errno == EINTR);
 
   if (fd < 0) {
     perror("failed to open /dev/urandom");
     abort();
   }
 
-  assert(kUnset == 0);
-  if (fd == kUnset) {
-    // Because we want to keep |urandom_fd| in the BSS, we have to initialise
-    // it to zero. But zero is a valid file descriptor too. Thus if open
-    // returns zero for /dev/urandom, we dup it to get a non-zero number.
-    fd = dup(fd);
-    close(kUnset);
-
-    if (fd <= 0) {
-      perror("failed to dup /dev/urandom fd");
-      abort();
-    }
-  }
-
   int flags = fcntl(fd, F_GETFD);
   if (flags == -1) {
     // Native Client doesn't implement |fcntl|.
@@ -342,40 +280,6 @@ static void wait_for_entropy(void) {
 #endif  // BORINGSSL_FIPS
 }
 
-void RAND_set_urandom_fd(int fd) {
-  fd = dup(fd);
-  if (fd < 0) {
-    perror("failed to dup supplied urandom fd");
-    abort();
-  }
-
-  assert(kUnset == 0);
-  if (fd == kUnset) {
-    // Because we want to keep |urandom_fd| in the BSS, we have to initialise
-    // it to zero. But zero is a valid file descriptor too. Thus if dup
-    // returned zero we dup it again to get a non-zero number.
-    fd = dup(fd);
-    close(kUnset);
-
-    if (fd <= 0) {
-      perror("failed to dup supplied urandom fd");
-      abort();
-    }
-  }
-
-  CRYPTO_STATIC_MUTEX_lock_write(rand_lock_bss_get());
-  *urandom_fd_requested_bss_get() = fd;
-  CRYPTO_STATIC_MUTEX_unlock_write(rand_lock_bss_get());
-
-  CRYPTO_once(rand_once_bss_get(), init_once);
-  if (*urandom_fd_bss_get() == kHaveGetrandom) {
-    close(fd);
-  } else if (*urandom_fd_bss_get() != fd) {
-    fprintf(stderr, "RAND_set_urandom_fd called after initialisation.\n");
-    abort();
-  }
-}
-
 // fill_with_entropy writes |len| bytes of entropy into |out|. It returns one
 // on success and zero on error. If |block| is one, this function will block
 // until the entropy pool is initialized. Otherwise, this function may fail,
@@ -397,7 +301,7 @@ static int fill_with_entropy(uint8_t *out, size_t len, int block, int seed) {
   }
 #endif
 
-  CRYPTO_once(rand_once_bss_get(), init_once);
+  CRYPTO_init_sysrand();
   if (block) {
     CRYPTO_once(wait_for_entropy_once_bss_get(), wait_for_entropy);
   }
@@ -452,30 +356,29 @@ void CRYPTO_sysrand(uint8_t *out, size_t requested) {
   }
 }
 
+void CRYPTO_init_sysrand(void) {
+  CRYPTO_once(rand_once_bss_get(), init_once);
+}
+
 #if defined(BORINGSSL_FIPS)
 void CRYPTO_sysrand_for_seed(uint8_t *out, size_t requested) {
   if (!fill_with_entropy(out, requested, /*block=*/1, /*seed=*/1)) {
     perror("entropy fill failed");
     abort();
   }
-
-#if defined(BORINGSSL_FIPS_BREAK_CRNG)
-  // This breaks the "continuous random number generator test" defined in FIPS
-  // 140-2, section 4.9.2, and implemented in rand_get_seed().
-  OPENSSL_memset(out, 0, requested);
-#endif
 }
+#endif  // BORINGSSL_FIPS
 
-void CRYPTO_sysrand_if_available(uint8_t *out, size_t requested) {
-  // Return all zeros if |fill_with_entropy| fails.
-  OPENSSL_memset(out, 0, requested);
-
-  if (!fill_with_entropy(out, requested, /*block=*/0, /*seed=*/0) &&
-      errno != EAGAIN) {
+int CRYPTO_sysrand_if_available(uint8_t *out, size_t requested) {
+  if (fill_with_entropy(out, requested, /*block=*/0, /*seed=*/0)) {
+    return 1;
+  } else if (errno == EAGAIN) {
+    OPENSSL_memset(out, 0, requested);
+    return 0;
+  } else {
     perror("opportunistic entropy fill failed");
     abort();
   }
 }
-#endif  // BORINGSSL_FIPS
 
 #endif  // OPENSSL_URANDOM

data/third_party/boringssl-with-bazel/src/crypto/fipsmodule/rsa/blinding.c

@@ -167,6 +167,10 @@ void BN_BLINDING_free(BN_BLINDING *r) {
   OPENSSL_free(r);
 }
 
+void BN_BLINDING_invalidate(BN_BLINDING *b) {
+  b->counter = BN_BLINDING_COUNTER - 1;
+}
+
 static int bn_blinding_update(BN_BLINDING *b, const BIGNUM *e,
                               const BN_MONT_CTX *mont, BN_CTX *ctx) {
   if (++b->counter == BN_BLINDING_COUNTER) {

data/third_party/boringssl-with-bazel/src/crypto/fipsmodule/rsa/internal.h

@@ -83,6 +83,7 @@ int rsa_default_private_transform(RSA *rsa, uint8_t *out, const uint8_t *in,
 
 BN_BLINDING *BN_BLINDING_new(void);
 void BN_BLINDING_free(BN_BLINDING *b);
+void BN_BLINDING_invalidate(BN_BLINDING *b);
 int BN_BLINDING_convert(BIGNUM *n, BN_BLINDING *b, const BIGNUM *e,
                         const BN_MONT_CTX *mont_ctx, BN_CTX *ctx);
 int BN_BLINDING_invert(BIGNUM *n, const BN_BLINDING *b, BN_MONT_CTX *mont_ctx,
@@ -107,6 +108,10 @@ int RSA_padding_check_PKCS1_OAEP_mgf1(uint8_t *out, size_t *out_len,
 int RSA_padding_add_none(uint8_t *to, size_t to_len, const uint8_t *from,
                          size_t from_len);
 
+// rsa_check_public_key checks that |rsa|'s public modulus and exponent are
+// within DoS bounds.
+int rsa_check_public_key(const RSA *rsa);
+
 // RSA_private_transform calls either the method-specific |private_transform|
 // function (if given) or the generic one. See the comment for
 // |private_transform| in |rsa_meth_st|.

data/third_party/boringssl-with-bazel/src/crypto/fipsmodule/rsa/rsa.c

@@ -167,6 +167,22 @@ int RSA_up_ref(RSA *rsa) {
 
 unsigned RSA_bits(const RSA *rsa) { return BN_num_bits(rsa->n); }
 
+const BIGNUM *RSA_get0_n(const RSA *rsa) { return rsa->n; }
+
+const BIGNUM *RSA_get0_e(const RSA *rsa) { return rsa->e; }
+
+const BIGNUM *RSA_get0_d(const RSA *rsa) { return rsa->d; }
+
+const BIGNUM *RSA_get0_p(const RSA *rsa) { return rsa->p; }
+
+const BIGNUM *RSA_get0_q(const RSA *rsa) { return rsa->q; }
+
+const BIGNUM *RSA_get0_dmp1(const RSA *rsa) { return rsa->dmp1; }
+
+const BIGNUM *RSA_get0_dmq1(const RSA *rsa) { return rsa->dmq1; }
+
+const BIGNUM *RSA_get0_iqmp(const RSA *rsa) { return rsa->iqmp; }
+
 void RSA_get0_key(const RSA *rsa, const BIGNUM **out_n, const BIGNUM **out_e,
                   const BIGNUM **out_d) {
   if (out_n != NULL) {
@@ -639,7 +655,15 @@ err:
 }
 
 static int check_mod_inverse(int *out_ok, const BIGNUM *a, const BIGNUM *ainv,
-                             const BIGNUM *m, int check_reduced, BN_CTX *ctx) {
+                             const BIGNUM *m, BN_CTX *ctx) {
+  if (BN_is_negative(ainv) || BN_cmp(ainv, m) >= 0) {
+    *out_ok = 0;
+    return 1;
+  }
+
+  // Note |bn_mul_consttime| and |bn_div_consttime| do not scale linearly, but
+  // checking |ainv| is in range bounds the running time, assuming |m|'s bounds
+  // were checked by the caller.
   BN_CTX_start(ctx);
   BIGNUM *tmp = BN_CTX_get(ctx);
   int ret = tmp != NULL &&
@@ -647,84 +671,97 @@ static int check_mod_inverse(int *out_ok, const BIGNUM *a, const BIGNUM *ainv,
             bn_div_consttime(NULL, tmp, tmp, m, ctx);
   if (ret) {
     *out_ok = BN_is_one(tmp);
-    if (check_reduced && (BN_is_negative(ainv) || BN_cmp(ainv, m) >= 0)) {
-      *out_ok = 0;
-    }
   }
   BN_CTX_end(ctx);
   return ret;
 }
 
 int RSA_check_key(const RSA *key) {
-  BIGNUM n, pm1, qm1, lcm, dmp1, dmq1, iqmp_times_q;
-  BN_CTX *ctx;
-  int ok = 0, has_crt_values;
+  // TODO(davidben): RSA key initialization is spread across
+  // |rsa_check_public_key|, |RSA_check_key|, |freeze_private_key|, and
+  // |BN_MONT_CTX_set_locked| as a result of API issues. See
+  // https://crbug.com/boringssl/316. As a result, we inconsistently check RSA
+  // invariants. We should fix this and integrate that logic.
 
   if (RSA_is_opaque(key)) {
     // Opaque keys can't be checked.
     return 1;
   }
 
+  if (!rsa_check_public_key(key)) {
+    return 0;
+  }
+
   if ((key->p != NULL) != (key->q != NULL)) {
     OPENSSL_PUT_ERROR(RSA, RSA_R_ONLY_ONE_OF_P_Q_GIVEN);
     return 0;
   }
 
-  if (!key->n || !key->e) {
-    OPENSSL_PUT_ERROR(RSA, RSA_R_VALUE_MISSING);
+  // |key->d| must be bounded by |key->n|. This ensures bounds on |RSA_bits|
+  // translate to bounds on the running time of private key operations.
+  if (key->d != NULL &&
+      (BN_is_negative(key->d) || BN_cmp(key->d, key->n) >= 0)) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_D_OUT_OF_RANGE);
     return 0;
   }
 
-  if (!key->d || !key->p) {
+  if (key->d == NULL || key->p == NULL) {
     // For a public key, or without p and q, there's nothing that can be
     // checked.
     return 1;
   }
 
-  ctx = BN_CTX_new();
+  BN_CTX *ctx = BN_CTX_new();
   if (ctx == NULL) {
     OPENSSL_PUT_ERROR(RSA, ERR_R_MALLOC_FAILURE);
     return 0;
   }
 
-  BN_init(&n);
+  BIGNUM tmp, de, pm1, qm1, dmp1, dmq1;
+  int ok = 0;
+  BN_init(&tmp);
+  BN_init(&de);
   BN_init(&pm1);
   BN_init(&qm1);
-  BN_init(&lcm);
   BN_init(&dmp1);
   BN_init(&dmq1);
-  BN_init(&iqmp_times_q);
 
-  int d_ok;
-  if (!bn_mul_consttime(&n, key->p, key->q, ctx) ||
-      // lcm = lcm(p, q)
-      !bn_usub_consttime(&pm1, key->p, BN_value_one()) ||
-      !bn_usub_consttime(&qm1, key->q, BN_value_one()) ||
-      !bn_lcm_consttime(&lcm, &pm1, &qm1, ctx) ||
-      // Other implementations use the Euler totient rather than the Carmichael
-      // totient, so allow unreduced |key->d|.
-      !check_mod_inverse(&d_ok, key->e, key->d, &lcm,
-                         0 /* don't require reduced */, ctx)) {
+  // Check that p * q == n. Before we multiply, we check that p and q are in
+  // bounds, to avoid a DoS vector in |bn_mul_consttime| below. Note that
+  // n was bound by |rsa_check_public_key|.
+  if (BN_is_negative(key->p) || BN_cmp(key->p, key->n) >= 0 ||
+      BN_is_negative(key->q) || BN_cmp(key->q, key->n) >= 0) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_N_NOT_EQUAL_P_Q);
+    goto out;
+  }
+  if (!bn_mul_consttime(&tmp, key->p, key->q, ctx)) {
     OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
     goto out;
   }
-
-  if (BN_cmp(&n, key->n) != 0) {
+  if (BN_cmp(&tmp, key->n) != 0) {
     OPENSSL_PUT_ERROR(RSA, RSA_R_N_NOT_EQUAL_P_Q);
     goto out;
   }
 
-  if (!d_ok) {
-    OPENSSL_PUT_ERROR(RSA, RSA_R_D_E_NOT_CONGRUENT_TO_1);
+  // d must be an inverse of e mod the Carmichael totient, lcm(p-1, q-1), but it
+  // may be unreduced because other implementations use the Euler totient. We
+  // simply check that d * e is one mod p-1 and mod q-1. Note d and e were bound
+  // by earlier checks in this function.
+  if (!bn_usub_consttime(&pm1, key->p, BN_value_one()) ||
+      !bn_usub_consttime(&qm1, key->q, BN_value_one()) ||
+      !bn_mul_consttime(&de, key->d, key->e, ctx) ||
+      !bn_div_consttime(NULL, &tmp, &de, &pm1, ctx) ||
+      !bn_div_consttime(NULL, &de, &de, &qm1, ctx)) {
+    OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
     goto out;
   }
 
-  if (BN_is_negative(key->d) || BN_cmp(key->d, key->n) >= 0) {
-    OPENSSL_PUT_ERROR(RSA, RSA_R_D_OUT_OF_RANGE);
+  if (!BN_is_one(&tmp) || !BN_is_one(&de)) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_D_E_NOT_CONGRUENT_TO_1);
     goto out;
   }
 
-  has_crt_values = key->dmp1 != NULL;
+  int has_crt_values = key->dmp1 != NULL;
   if (has_crt_values != (key->dmq1 != NULL) ||
       has_crt_values != (key->iqmp != NULL)) {
     OPENSSL_PUT_ERROR(RSA, RSA_R_INCONSISTENT_SET_OF_CRT_VALUES);
@@ -733,12 +770,9 @@ int RSA_check_key(const RSA *key) {
 
   if (has_crt_values) {
     int dmp1_ok, dmq1_ok, iqmp_ok;
-    if (!check_mod_inverse(&dmp1_ok, key->e, key->dmp1, &pm1,
-                           1 /* check reduced */, ctx) ||
-        !check_mod_inverse(&dmq1_ok, key->e, key->dmq1, &qm1,
-                           1 /* check reduced */, ctx) ||
-        !check_mod_inverse(&iqmp_ok, key->q, key->iqmp, key->p,
-                           1 /* check reduced */, ctx)) {
+    if (!check_mod_inverse(&dmp1_ok, key->e, key->dmp1, &pm1, ctx) ||
+        !check_mod_inverse(&dmq1_ok, key->e, key->dmq1, &qm1, ctx) ||
+        !check_mod_inverse(&iqmp_ok, key->q, key->iqmp, key->p, ctx)) {
       OPENSSL_PUT_ERROR(RSA, ERR_LIB_BN);
       goto out;
     }
@@ -752,13 +786,12 @@ int RSA_check_key(const RSA *key) {
   ok = 1;
 
 out:
-  BN_free(&n);
+  BN_free(&tmp);
+  BN_free(&de);
   BN_free(&pm1);
   BN_free(&qm1);
-  BN_free(&lcm);
   BN_free(&dmp1);
   BN_free(&dmq1);
-  BN_free(&iqmp_times_q);
   BN_CTX_free(ctx);
 
   return ok;

data/third_party/boringssl-with-bazel/src/crypto/fipsmodule/rsa/rsa_impl.c

@@ -70,9 +70,15 @@
 #include "../bn/internal.h"
 #include "../../internal.h"
 #include "../delocate.h"
+#include "../rand/fork_detect.h"
 
 
-static int check_modulus_and_exponent_sizes(const RSA *rsa) {
+int rsa_check_public_key(const RSA *rsa) {
+  if (rsa->n == NULL || rsa->e == NULL) {
+    OPENSSL_PUT_ERROR(RSA, RSA_R_VALUE_MISSING);
+    return 0;
+  }
+
   unsigned rsa_bits = BN_num_bits(rsa->n);
 
   if (rsa_bits > 16 * 1024) {
@@ -252,8 +258,7 @@ size_t rsa_default_size(const RSA *rsa) {
 
 int RSA_encrypt(RSA *rsa, size_t *out_len, uint8_t *out, size_t max_out,
                 const uint8_t *in, size_t in_len, int padding) {
-  if (rsa->n == NULL || rsa->e == NULL) {
-    OPENSSL_PUT_ERROR(RSA, RSA_R_VALUE_MISSING);
+  if (!rsa_check_public_key(rsa)) {
     return 0;
   }
 
@@ -268,10 +273,6 @@ int RSA_encrypt(RSA *rsa, size_t *out_len, uint8_t *out, size_t max_out,
     return 0;
   }
 
-  if (!check_modulus_and_exponent_sizes(rsa)) {
-    return 0;
-  }
-
   ctx = BN_CTX_new();
   if (ctx == NULL) {
     goto err;
@@ -345,7 +346,7 @@ err:
 // MAX_BLINDINGS_PER_RSA defines the maximum number of cached BN_BLINDINGs per
 // RSA*. Then this limit is exceeded, BN_BLINDING objects will be created and
 // destroyed as needed.
-#if defined(OPNESSL_TSAN)
+#if defined(OPENSSL_TSAN)
 // Smaller under TSAN so that the edge case can be hit with fewer threads.
 #define MAX_BLINDINGS_PER_RSA 2
 #else
@@ -365,8 +366,21 @@ static BN_BLINDING *rsa_blinding_get(RSA *rsa, unsigned *index_used,
   assert(rsa->mont_n != NULL);
 
   BN_BLINDING *ret = NULL;
+  const uint64_t fork_generation = CRYPTO_get_fork_generation();
   CRYPTO_MUTEX_lock_write(&rsa->lock);
 
+  // Wipe the blinding cache on |fork|.
+  if (rsa->blinding_fork_generation != fork_generation) {
+    for (unsigned i = 0; i < rsa->num_blindings; i++) {
+      // The inuse flag must be zero unless we were forked from a
+      // multi-threaded process, in which case calling back into BoringSSL is
+      // forbidden.
+      assert(rsa->blindings_inuse[i] == 0);
+      BN_BLINDING_invalidate(rsa->blindings[i]);
+    }
+    rsa->blinding_fork_generation = fork_generation;
+  }
+
   uint8_t *const free_inuse_flag =
       OPENSSL_memchr(rsa->blindings_inuse, 0, rsa->num_blindings);
   if (free_inuse_flag != NULL) {
@@ -578,8 +592,7 @@ static int mod_exp(BIGNUM *r0, const BIGNUM *I, RSA *rsa, BN_CTX *ctx);
 
 int RSA_verify_raw(RSA *rsa, size_t *out_len, uint8_t *out, size_t max_out,
                    const uint8_t *in, size_t in_len, int padding) {
-  if (rsa->n == NULL || rsa->e == NULL) {
-    OPENSSL_PUT_ERROR(RSA, RSA_R_VALUE_MISSING);
+  if (!rsa_check_public_key(rsa)) {
     return 0;
   }
 
@@ -596,10 +609,6 @@ int RSA_verify_raw(RSA *rsa, size_t *out_len, uint8_t *out, size_t max_out,
     return 0;
   }
 
-  if (!check_modulus_and_exponent_sizes(rsa)) {
-    return 0;
-  }
-
   BN_CTX *ctx = BN_CTX_new();
   if (ctx == NULL) {
     return 0;
@@ -924,20 +933,57 @@ static int ensure_bignum(BIGNUM **out) {
   return *out != NULL;
 }
 
-// kBoringSSLRSASqrtTwo is the BIGNUM representation of ⌊2¹⁵³⁵×√2⌋. This is
-// chosen to give enough precision for 3072-bit RSA, the largest key size FIPS
+// kBoringSSLRSASqrtTwo is the BIGNUM representation of ⌊2²⁰⁴⁷×√2⌋. This is
+// chosen to give enough precision for 4096-bit RSA, the largest key size FIPS
 // specifies. Key sizes beyond this will round up.
 //
-// To verify this number, check that n² < 2³⁰⁷¹ < (n+1)², where n is value
+// To calculate, use the following Haskell code:
+//
+// import Text.Printf (printf)
+// import Data.List (intercalate)
+//
+// pow2 = 4095
+// target = 2^pow2
+//
+// f x = x*x - (toRational target)
+//
+// fprime x = 2*x
+//
+// newtonIteration x = x - (f x) / (fprime x)
+//
+// converge x =
+//   let n = floor x in
+//   if n*n - target < 0 && (n+1)*(n+1) - target > 0
+//     then n
+//     else converge (newtonIteration x)
+//
+// divrem bits x = (x `div` (2^bits), x `rem` (2^bits))
+//
+// bnWords :: Integer -> [Integer]
+// bnWords x =
+//   if x == 0
+//     then []
+//     else let (high, low) = divrem 64 x in low : bnWords high
+//
+// showWord x = let (high, low) = divrem 32 x in printf "TOBN(0x%08x, 0x%08x)" high low
+//
+// output :: String
+// output = intercalate ", " $ map showWord $ bnWords $ converge (2 ^ (pow2 `div` 2))
+//
+// To verify this number, check that n² < 2⁴⁰⁹⁵ < (n+1)², where n is value
 // represented here. Note the components are listed in little-endian order. Here
 // is some sample Python code to check:
 //
 //   >>> TOBN = lambda a, b: a << 32 | b
 //   >>> l = [ <paste the contents of kSqrtTwo> ]
 //   >>> n = sum(a * 2**(64*i) for i, a in enumerate(l))
-//   >>> n**2 < 2**3071 < (n+1)**2
+//   >>> n**2 < 2**4095 < (n+1)**2
 //   True
 const BN_ULONG kBoringSSLRSASqrtTwo[] = {
+    TOBN(0x4d7c60a5, 0xe633e3e1), TOBN(0x5fcf8f7b, 0xca3ea33b),
+    TOBN(0xc246785e, 0x92957023), TOBN(0xf9acce41, 0x797f2805),
+    TOBN(0xfdfe170f, 0xd3b1f780), TOBN(0xd24f4a76, 0x3facb882),
+    TOBN(0x18838a2e, 0xaff5f3b2), TOBN(0xc1fcbdde, 0xa2f7dc33),
     TOBN(0xdea06241, 0xf7aa81c2), TOBN(0xf6a1be3f, 0xca221307),
     TOBN(0x332a5e9f, 0x7bda1ebf), TOBN(0x0104dc01, 0xfe32352f),
     TOBN(0xb8cf341b, 0x6f8236c7), TOBN(0x4264dabc, 0xd528b651),
@@ -1107,8 +1153,8 @@ static int rsa_generate_key_impl(RSA *rsa, int bits, const BIGNUM *e_value,
 
   // Reject excessively large public exponents. Windows CryptoAPI and Go don't
   // support values larger than 32 bits, so match their limits for generating
-  // keys. (|check_modulus_and_exponent_sizes| uses a slightly more conservative
-  // value, but we don't need to support generating such keys.)
+  // keys. (|rsa_check_public_key| uses a slightly more conservative value, but
+  // we don't need to support generating such keys.)
   // https://github.com/golang/go/issues/3161
   // https://msdn.microsoft.com/en-us/library/aa387685(VS.85).aspx
   if (BN_num_bits(e_value) > 32) {
@@ -1158,13 +1204,13 @@ static int rsa_generate_key_impl(RSA *rsa, int bits, const BIGNUM *e_value,
   int sqrt2_bits = kBoringSSLRSASqrtTwoLen * BN_BITS2;
   assert(sqrt2_bits == (int)BN_num_bits(sqrt2));
   if (sqrt2_bits > prime_bits) {
-    // For key sizes up to 3072 (prime_bits = 1536), this is exactly
+    // For key sizes up to 4096 (prime_bits = 2048), this is exactly
     // ⌊2^(prime_bits-1)×√2⌋.
     if (!BN_rshift(sqrt2, sqrt2, sqrt2_bits - prime_bits)) {
       goto bn_err;
     }
   } else if (prime_bits > sqrt2_bits) {
-    // For key sizes beyond 3072, this is approximate. We err towards retrying
+    // For key sizes beyond 4096, this is approximate. We err towards retrying
     // to ensure our key is the right size and round up.
     if (!BN_add_word(sqrt2, 1) ||
         !BN_lshift(sqrt2, sqrt2, prime_bits - sqrt2_bits)) {
@@ -1270,58 +1316,80 @@ static void replace_bn_mont_ctx(BN_MONT_CTX **out, BN_MONT_CTX **in) {
   *in = NULL;
 }
 
-int RSA_generate_key_ex(RSA *rsa, int bits, const BIGNUM *e_value,
-                        BN_GENCB *cb) {
+static int RSA_generate_key_ex_maybe_fips(RSA *rsa, int bits,
+                                          const BIGNUM *e_value, BN_GENCB *cb,
+                                          int check_fips) {
+  RSA *tmp = NULL;
+  uint32_t err;
+  int ret = 0;
+
   // |rsa_generate_key_impl|'s 2^-20 failure probability is too high at scale,
   // so we run the FIPS algorithm four times, bringing it down to 2^-80. We
   // should just adjust the retry limit, but FIPS 186-4 prescribes that value
   // and thus results in unnecessary complexity.
-  for (int i = 0; i < 4; i++) {
+  int failures = 0;
+  do {
     ERR_clear_error();
     // Generate into scratch space, to avoid leaving partial work on failure.
-    RSA *tmp = RSA_new();
+    tmp = RSA_new();
     if (tmp == NULL) {
-      return 0;
+      goto out;
     }
+
     if (rsa_generate_key_impl(tmp, bits, e_value, cb)) {
-      replace_bignum(&rsa->n, &tmp->n);
-      replace_bignum(&rsa->e, &tmp->e);
-      replace_bignum(&rsa->d, &tmp->d);
-      replace_bignum(&rsa->p, &tmp->p);
-      replace_bignum(&rsa->q, &tmp->q);
-      replace_bignum(&rsa->dmp1, &tmp->dmp1);
-      replace_bignum(&rsa->dmq1, &tmp->dmq1);
-      replace_bignum(&rsa->iqmp, &tmp->iqmp);
-      replace_bn_mont_ctx(&rsa->mont_n, &tmp->mont_n);
-      replace_bn_mont_ctx(&rsa->mont_p, &tmp->mont_p);
-      replace_bn_mont_ctx(&rsa->mont_q, &tmp->mont_q);
-      replace_bignum(&rsa->d_fixed, &tmp->d_fixed);
-      replace_bignum(&rsa->dmp1_fixed, &tmp->dmp1_fixed);
-      replace_bignum(&rsa->dmq1_fixed, &tmp->dmq1_fixed);
-      replace_bignum(&rsa->inv_small_mod_large_mont,
-                     &tmp->inv_small_mod_large_mont);
-      rsa->private_key_frozen = tmp->private_key_frozen;
-      RSA_free(tmp);
-      return 1;
+      break;
     }
-    uint32_t err = ERR_peek_error();
+
+    err = ERR_peek_error();
     RSA_free(tmp);
     tmp = NULL;
+    failures++;
+
     // Only retry on |RSA_R_TOO_MANY_ITERATIONS|. This is so a caller-induced
     // failure in |BN_GENCB_call| is still fatal.
-    if (ERR_GET_LIB(err) != ERR_LIB_RSA ||
-        ERR_GET_REASON(err) != RSA_R_TOO_MANY_ITERATIONS) {
-      return 0;
-    }
+  } while (failures < 4 && ERR_GET_LIB(err) == ERR_LIB_RSA &&
+           ERR_GET_REASON(err) == RSA_R_TOO_MANY_ITERATIONS);
+
+  if (tmp == NULL || (check_fips && !RSA_check_fips(tmp))) {
+    goto out;
   }
 
-  return 0;
+  replace_bignum(&rsa->n, &tmp->n);
+  replace_bignum(&rsa->e, &tmp->e);
+  replace_bignum(&rsa->d, &tmp->d);
+  replace_bignum(&rsa->p, &tmp->p);
+  replace_bignum(&rsa->q, &tmp->q);
+  replace_bignum(&rsa->dmp1, &tmp->dmp1);
+  replace_bignum(&rsa->dmq1, &tmp->dmq1);
+  replace_bignum(&rsa->iqmp, &tmp->iqmp);
+  replace_bn_mont_ctx(&rsa->mont_n, &tmp->mont_n);
+  replace_bn_mont_ctx(&rsa->mont_p, &tmp->mont_p);
+  replace_bn_mont_ctx(&rsa->mont_q, &tmp->mont_q);
+  replace_bignum(&rsa->d_fixed, &tmp->d_fixed);
+  replace_bignum(&rsa->dmp1_fixed, &tmp->dmp1_fixed);
+  replace_bignum(&rsa->dmq1_fixed, &tmp->dmq1_fixed);
+  replace_bignum(&rsa->inv_small_mod_large_mont,
+                 &tmp->inv_small_mod_large_mont);
+  rsa->private_key_frozen = tmp->private_key_frozen;
+  ret = 1;
+
+out:
+  RSA_free(tmp);
+  return ret;
+}
+
+int RSA_generate_key_ex(RSA *rsa, int bits, const BIGNUM *e_value,
+                        BN_GENCB *cb) {
+  return RSA_generate_key_ex_maybe_fips(rsa, bits, e_value, cb,
+                                        /*check_fips=*/0);
 }
 
 int RSA_generate_key_fips(RSA *rsa, int bits, BN_GENCB *cb) {
   // FIPS 186-4 allows 2048-bit and 3072-bit RSA keys (1024-bit and 1536-bit
   // primes, respectively) with the prime generation method we use.
-  if (bits != 2048 && bits != 3072) {
+  // Subsequently, IG A.14 stated that larger modulus sizes can be used and ACVP
+  // testing supports 4096 bits.
+  if (bits != 2048 && bits != 3072 && bits != 4096) {
     OPENSSL_PUT_ERROR(RSA, RSA_R_BAD_RSA_PARAMETERS);
     return 0;
   }
@@ -1329,8 +1397,7 @@ int RSA_generate_key_fips(RSA *rsa, int bits, BN_GENCB *cb) {
   BIGNUM *e = BN_new();
   int ret = e != NULL &&
             BN_set_word(e, RSA_F4) &&
-            RSA_generate_key_ex(rsa, bits, e, cb) &&
-            RSA_check_fips(rsa);
+            RSA_generate_key_ex_maybe_fips(rsa, bits, e, cb, /*check_fips=*/1);
   BN_free(e);
   return ret;
 }